Research On Visual Optimization Algorithms Based On Deep Learning

About 80% of the information resources in human life come from video images,and video images have become an important carrier of human information transmission.However,in the process of obtaining video images by using cameras and other photographic equipment,the final presentation quality of video images is reduced due to the influence of the equipment itself and external factors.Therefore,the generated video images need to be further optimized.This paper focuses on two visual optimization problems: removing image motion blur and improving video frame rate.The traditional image motion deblurring technology can achieve good deblurring effect under general and certain blur type.However,in complex motion scenes,due to the complexity and diversity of blur types,and affected by noise and inaccurate estimation of point spread function,there will still be blurry in video images after deblurring,the result is not satisfactory.In addition,the improvement of video frame rate depends on the synthesis of intermediate frames,which requires the study of video interpolation algorithm.The pixel synthesis effect of traditional video interpolation technologies mainly depends on the accuracy of optical flow estimation,which is difficult to ensure its accuracy in complex scenes such as large motion displacement,sudden changes in lighting conditions,and occlusion of images.In recent years,with the development of data science and the progress of computer science,using deep learning technology to solve the relevant problems in the field of computer vision has become an important direction of scientific research.At present,researcher scholars have proposed some methods based on neural network models in the research direction of image motion deblurring and video interpolation,and achieved very good results.On the basis of previous studies,this paper further explores and innovates the application of deep learning in the research direction of image motion deblurring and video interpolation,and proposes two corresponding deep learning algorithms,the specific work is as follows:(1)Image motion deblurring method based on conditional generative adversarial network.The core task of this method is to construct two subnetworks in conditional generative adversarial network.The generative network adopts a designed SAD-Dense Net neural network,the long and short jump connection and information splicing technology are used in shallow feature extraction part and deep feature extraction part for many times,which can effectively solve the problem of low utilization rate of image feature information.The adversarial network uses a special convolutional neural network to discriminate the authenticity of images.This method can comprehensively enhance the learning of high-level features in the image data and effectively restore the motion blurred image to a clear image.(2)Video interpolation method based on compression and refined deep voxel flow neural network.This method widens the deep voxel flow network horizontally,designs the information supplement network structure from the coarse voxel flow to the fine voxel flow,and uses the difference between the input frame and the coarse interpolation frame to supplement the edge details to improve the interpolation accuracy.In order to keep the lightness of network parameters,the number of redundant convolutional network channels is removed by parameter compression technology for reducing the increase of parameters.This method improves the insufficiency of the deep voxel flow network in detail,improves the quality of interpolation frame,and does not significantly increase the number of parameters.The two methods mentioned above are proved to be feasible through experiments.These two methods both adopt an end-to-end approach,and the time complexity of training the model is not high,which can achieve high-efficiency and high-quality processing of video images.